A basic outline of the steps I take to go from raw data to deployed model.

data-science-process.md

Data Science Process

Steps to take for a comprehensive analysis.

Stage 1: Define

Project background.

Step 1.1: Describe.

Define project goals and objectives.

• What is the project about?
• What is the goal of this project?

Step 1.2: Deter.

Consider potential setbacks.

• What are some areas of trouble?
• Are there any disadvantages to the models you're planning on building?
• Do you anticipate anything to be wrong with the data?
• What precautions should be taken?

Stage 2: Data

Manipulation and analysis.

Step 2.1: Download.

Collect and store the data.

• Where is the data coming from?
• How did you get it?
• Did you import the data successfully?
• Can you access it correctly?
  • check the head and tail

Step 2.2: Distill.

Preprocess and tidy data.

1. Take a look through the dataset.

• check the dimensions, shape
• look at the structure and summary statistics
• use this source for inspiration

2. Check for bad or irrelevant data.

• remove any duplicate data points
• minimize the dataset by choosing observations beneficial to analysis
• drop any features that will prove inconsequential to analysis or modeling

3. Check for missing data.

• drop any egregiously unavailable observations
• note any remaining missing data

4. Check for mislabeled or inconsistent data.

• clean and combine variations on the same response
• rename column headers if necessary
• create or consolidate columns where helpful

Step 2.3: Discover.

Perform exploratory analysis.

1. What does the data look like now?

• check the dimensions, shape
• look at the structure and summary statistics

2. What kind of data are you working with?

• use this resource
• check for useless, nominal, binary, ordinal, count, time, interval, + image, video, audio, text
• will the data need any transformation at this point?

3. What's the distribution of the data?

• use histograms for a visual check
• use plots to confirm
• do you need to normalize the data for your model(s)?

4. Are there any correlated variables?

• use correlation plot for visual check
• use correlation matrix to confirm
• do you need to remove these relationships?

5. Are there any outliers?

• use boxplots for a visual check
• use hypothesis tests to confirm
• do you need to remove these outliers? look at them more closely?

6. Is there any missing data?

• check for NULL values
• use the missingno package to visualize
• do you need to remove these observations?
• can you impute the missing data? should you?
• what affect will this have on your model(s)?

7. Summarize your findings so far.

• what does the data mean?
• what kind of data do you have?
• what transformations, if any, are needed?
• what can you see with this data? what CAN'T you see?
• can this data help you achieve the objectives outlined in Part 1?

Step 2.4: Dissect.

Transform the dataset.

• Perform any necessary transformations before building models.
  • encoding, imputation, outlier removal
  • normalization, scaling, centering
  • dimensionality reduction
• Explain why certain transformations were not done.
• Explain if the data set needs no transformation.

Step 2.5: Divide.

Split the data for modeling.

• Split the data into train/validation/test sets.

Stage 3: Develop

Modeling and prediction.

Step 3.1: Deliberate.

Train on a few different models.

• Identify candidate models.

  • what kind of model(s) do you need to achieve the Part 1 goal?
  • what kind of model(s) work best with the data you have on hand?
  • will you build your own models or use available ones?
  • how will you measure the quality of your chosen model(s)?

• Build preliminary models on the training set.

• Store preliminary measures of model quality.

  • use a cross-model validation measure, if possible

Step 3.2: Decide.

Validate to choose the best model.

• Identify final model(s).
  • how did the model(s) perform on the training set?
• Check preliminary models on the validation set.
• Store intermediary measures of model quality.
  • may need to use cross-validation on multiple models

Step 3.3: Declare.

Test to evaluate model performance.

• Test chosen model(s).
  • how well does the model perform on unseen data?
• Check chosen models on the test set.
• Store final measures of model quality.

Stage 4: Deploy

Production and distribution.

Step 4.1: Demonstrate.

Demonstrate model performance.

• Make predictions on unseen data.

Step 4.2: Distribute.

Deploy model to production.

• Make model available to other users.

Stage 5: Discuss

Recap and reflection.

Step 5.1: Determine.

Review the final model.

• What conclusion(s) did you reach?
• How did your results align with or differ from your expectations?

Step 5.2: Discourage.

Discuss challenges and obstacles.

• Did you run into any major issues?
• Did you run into any minor issues?
• Were there any bugs to work through?
• Was there anything you just couldn't solve this go round?

Step 5.3: Direct.

Opportunities for future research.

• What could you do differently next time?
• How could you extend this project?
• What would you like to try next?

Step 5.4: Disseminate.

Resources for more information.

• How can we access resources used in this analysis?
• Where can readers find more information?
  • "The following resources were of immense help to me as I completed this project:"